predict-addict.bsky.social
PhD in machine learning | conformal prediction | time-series | author of bestselling Practical Guide to Applied Conformal-Prediction https://a.co/d/iHRag4i
6,161 posts
206 followers
30 following
Regular Contributor
Active Commenter
comment in response to
post
www.kevinsheppard.co...
comment in response to
post
iclr-blogposts.githu...
comment in response to
post
That critique, originally in a tweet, just made it into both the academic discussion and the official blogposts of the leading ML conference.
The line?
"When it comes to TabPFN, F stands for 'Frankenstein.'" (c)
Turns out, calling out hype with reason sometimes echoes where it matters most.
comment in response to
post
Back in 2023, when TabPFN was parading itself as the one-model-to-rule-them-all, I — along a few others — raised concerns about its synthetic training data and real-world reliability.
comment in response to
post
Like conformal prediction for example.
#research #conformalprediction
comment in response to
post
In fact, a recent paper once again confirms CatBoost's dominance with tabular data, while XGBoost came in at just … number 10.
“AComprehensive Benchmark of Machine and Deep Learning Across Diverse Tabular Datasets”
#tabulardata #catboost
comment in response to
post
It's time to set the record straight: XGBoost might get a lot of attention on social media, but when it comes to handling tabular data, CatBoost consistently outperforms.
Study after study has shown that CatBoost is the leader in this space.
comment in response to
post
#forecasting #timeseries #data #analytics #machinelearning #modernforecasting #edarealness #modelingmindset #businessimpact
comment in response to
post
📈 This is the deepest course on forecasting out there—no fluff, all signal. Companies are seeing instant ROI.
🚀 Enrollment is now open for Cohort 2—join before the next price hike hits soon.
maven.com/valeriy-ma...
comment in response to
post
That’s why I’m beyond proud of the incredible students in my first completely oversubscribed cohort of Modern Time Series Forecasting. They're not just learning tools—they're learning the kata of forecasting.
The mindset, the principles, the diagnostics. They're already delivering real impact.
comment in response to
post
❌ Skip EDA → Miss trends, seasonality, anomalies
❌ Ignore stationarity → Faulty assumptions
❌ No benchmarks → No way to measure value
❌ Blindly trust LSTM/Prophet/LLMs → Bad forecasts
Forecasting is a discipline—built on fundamentals, not fads.
comment in response to
post
in critical applications like human health, finance and self driving cars? no reasonable person will.
arxiv.org/pdf/2502.0...
#tabulardata
comment in response to
post
Until then, let’s call this what it is: intellectual laziness masquerading as empirical rigor.
📄 Read the paper if you must
#TimeSeries #MachineLearning #BadScience #SyntheticData #AISnakeOil #GPDelusions #MLResearch
comment in response to
post
This kind of work misleads the field, wastes researcher time, and creates the illusion of progress where there is none.
If you want to build meaningful time series models:
Benchmark on real datasets.
Embrace messiness.
Stop gaming fake metrics on fake data.
comment in response to
post
Zero external validity: Not a single claim about performance under uncertainty, regime shifts, or multivariate noise holds water when your entire benchmark is generated by kernel magic and white noise.
🚨 This is not science. It’s curve-fitting on fantasy.
comment in response to
post
Architectural theater: Designing and "tailoring" model architectures to perform well on simulated GP samples is like tuning a race car for a treadmill. You’re optimizing for a cartoon version of reality.
comment in response to
post
Synthetic delusion: Gaussian processes are mathematically cute—but they bear no resemblance to the chaotic, non-stationary, heteroskedastic mess that defines real-world time series.
comment in response to
post
The 'idea of a century' : Let’s test forecasting models… on fake, smooth, stationary, Gaussian toy data... and then draw bold conclusions about architecture design.
This is not just misguided—it’s intellectually bankrupt.
🔥 Why this is a new low:
comment in response to
post
If you’re serious about AI, quantum computing, or advanced math, this is your holy grail. 📖✨
Drop a ♥️ if you want the link! (And brace yourself—this is next-level knowledge.)
#AI #MachineLearning #MathGenius #HiddenGem #TheMatrixButReal
comment in response to
post
💎 Why this blows everything else out of the water:
Pure, unfiltered depth—no fluff, just the foundational truths of linear algebra.
Criminally underrated—most don’t even know it exists, but the pros swear by it.
Volume 1? Buckle up—this is just the beginning.
comment in response to
post
The golden age of being paid a premium just for writing code is over—unless you’re at the top of your game.
comment in response to
post
company said so.
So if anyone today is spending their time grinding LeetCode or SQL hoping it’ll guarantee a career, they’re performing the intellectual equivalent of seppuku—slowly killing their prospects instead of focusing on what actually matters.
It’s time to get real.
comment in response to
post
Worse still, with over a billion people globally ready to work for any wage, it makes zero sense for someone starting out in 2025 to pour months into LeetCode grinding or spend thousands on glorified SQL courses—just because a guy with a slick LinkedIn profile and a background at a "cool" tech
comment in response to
post
And here’s the harsh economic truth: when barriers to entry disappear, so do premiums.
Average or even mediocre coders could once grind LeetCode, memorize patterns, and land six-figure jobs at Google or Meta. That era is closing fast. You can’t fake it anymore.
comment in response to
post
We even stopped calling ourselves “programmers” and rebranded as “devs.” That was the first crack in the wall.
The second—and far more disruptive—blow has been the rise of LLMs that can now write code. This isn't a passing phase. It’s here to stay.
comment in response to
post
The early programmers—the ones who learned C from K&R matrix printer printouts—thrived because they were operating in a space few could even enter. They were elite.
Then came Google, Stack Overflow, and the rise of the internet. Suddenly, programming became more accessible. The mystique faded.
comment in response to
post
The truth is, some people simply aren’t built to adapt. Remember Darwin?
Let me offer my perspective as someone who’s been coding since school. The last 20–30 years of comfort and insulation in the software industry have been an aberration.
comment in response to
post
Feliks Gantmacher’s name might not trend on social media, but his fingerprints are everywhere — in linear systems, missile guidance, mechanical theory, and even the DNA of early AI.
A mathematician. An engineer. A mentor.
A true warrior-scholar.
archive.org/details/...
#physics #ai #math
comment in response to
post
Gantmacher’s expertise in systems, feedback, and stability naturally fed into the mathematical foundations of what we now call AI. So while he may not have called it that, his work helped build it.
🏛️ Legacy
comment in response to
post
Though not formally labeled an AI researcher, Gantmacher was closely associated with Mark Aizerman, a pivotal figure in Soviet cybernetics and early machine learning.
comment in response to
post
As department chair at MIPT (Moscow Institute of Physics and Technology), Gantmacher mentored a generation of scientists, engineers, and mathematicians. His influence shaped Soviet education and research for decades.
🤖 Close to the Roots of AI
comment in response to
post
Just 262 pages long, this concise volume is an absolute gem: clear, rigorous, and elegant. It was once one of my favorite books — a perfect bridge between theoretical physics and applied mathematics.
🧬 Builder of Scientific Legacy
comment in response to
post
To this day, they’re cited not only for their mathematical depth but for their lucid exposition.
📖 A Hidden Gem: Lectures on Mechanics
And luckily for English-speaking readers, another of Gantmacher’s masterpieces — Lectures in Analytical Mechanics — is available in translation!
comment in response to
post
Post-war, Gantmacher penned one of the most influential works in applied mathematics: "Theory of Matrices" (Vols. 1 & 2). These volumes are still foundational today — from control theory to quantum mechanics.
comment in response to
post
His mastery of stability theory, vibration analysis, and control systems was applied directly to Soviet military technology, earning him a reputation as a “warrior mathematician.”
📚 Mastermind of Matrix Theory
comment in response to
post
During WWII, Gantmacher was no armchair mathematician. He worked on the development of the legendary MLRS “Katyusha”, contributing to one of the most iconic rocket artillery systems in history.
comment in response to
post
Few figures in 20th-century science straddled as many worlds as Feliks Ruvimovich Gantmacher — a man whose legacy stretches from the thunder of WWII rocket launchers to the elegance of matrix algebra and the quiet rigor of MIPT lecture halls.
🚀 From War Rooms to Weapon Systems
comment in response to
post
Early readers believe Mastering Modern Time Series Forecasting is well on its way to becoming a definitive resource in the field—praised for its depth, clarity, and relevance.
valeman.gumroad.com/...
#timeseries #forecasting